ML pipeline best practices: reproducibility as a first-class citizen
2026-03-28 · Sofia Lindqvist
Seven principles we follow when designing pipelines that survive contact with production traffic, team changes, and two-year audits.
Blog
Engineering, ML, and product updates.
Model monitoring in production: what to actually watch
2026-03-14 · Mikael Laakso
A practical guide to which ML monitoring signals matter, which are noise, and what to do when one of them fires at 2am.
Shadow deployments for ML: catching regressions before they hit users
2026-02-22 · Mikael Laakso
How we use shadow traffic to validate new models on real production data without risking a single user-visible error.
Reproducibility in ML: why your training runs don't match and how to fix it
2026-02-08 · Sofia Lindqvist
Five sources of non-determinism in typical ML pipelines and the practical controls that make training runs reproducible.
Data drift detection: PSI vs Kolmogorov–Smirnov in practice
2026-01-20 · Emma Schmidt
We compared PSI and KS test on six months of production features. Here's what each is good at, and when to prefer one over the other.
GPU scheduling on Kubernetes: bin-packing for ML workloads
2025-12-12 · Jan van der Berg
How we cut GPU spend by 38% with a smarter bin-packing scheduler and fractional GPU allocation.
MLOps stacks in 2025: a comparison of five approaches we evaluated
2025-11-18 · Sofia Lindqvist
From monolithic platforms to best-of-breed stacks, here's what we found when our team evaluated five MLOps architectures for a mid-size ML org.
Feature store: build vs buy, six months in
2025-10-25 · Mikael Laakso
We tried both. Here's what we learned from six months of running a self-built feature store before switching to a managed one.